Skip to main content

OpenVLA — The Open-Source VLA Ecosystem

Duration: 60 min · Level: Advanced · Module: 5. Foundation Models & VLA Architecture · Focus: OpenVLA, open-source, fine-tuning, VLA

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • OpenVLA architecture
  • Performance
  • Fine-tuning
  • Action tokenization
  • OpenVLA-OFT (2025 follow-up)

Why this matters​

OpenVLA (June 2024, Stanford + Berkeley) is the first fully open-source VLA that matches RT-2-X performance on standard benchmarks.

Overview​

OpenVLA (June 2024, Stanford + Berkeley) is the first fully open-source VLA that matches RT-2-X performance on standard benchmarks. Its release democratized VLA research: any lab can fine-tune a state-of-the-art general-purpose robot policy on their own hardware.

Key concepts​

Key idea

OpenVLA architecture: 7.5B parameters based on Prismatic VLM (SigLIP vision encoder + Llama 2 language model); actions tokenized as discrete language tokens

  • Performance: matches RT-2-X on BridgeV2 benchmark, outperforms on several manipulation tasks; runs at 6 Hz on single A100 GPU, 1.5 Hz on NVIDIA Orin AGX
  • Fine-tuning: LoRA fine-tuning on 200-500 demonstrations converges in 2-4 hours on single GPU; makes task-specific adaptation practical for small labs
  • Action tokenization: continuous actions discretized to 256 bins per dimension, encoded as language tokens; allows use of LLM training infrastructure directly
  • OpenVLA-OFT (2025 follow-up): parallel decoding + action chunking reduces latency to 25+ Hz; enables contact-rich, high-frequency manipulation
  • Available at: github.com/openvla/openvla with pre-trained checkpoints on HuggingFace; Apache 2.0 license

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about OpenVLA architecture?

7.5B parameters based on Prismatic VLM (SigLIP vision encoder + Llama 2 language model); actions tokenized as discrete language tokens

Q2. What do you know about Performance?

matches RT-2-X on BridgeV2 benchmark, outperforms on several manipulation tasks; runs at 6 Hz on single A100 GPU, 1.5 Hz on NVIDIA Orin AGX

Q3. What do you know about Fine-tuning?

LoRA fine-tuning on 200-500 demonstrations converges in 2-4 hours on single GPU; makes task-specific adaptation practical for small labs

Q4. What do you know about Action tokenization?

continuous actions discretized to 256 bins per dimension, encoded as language tokens; allows use of LLM training infrastructure directly

Q5. What do you know about OpenVLA-OFT (2025 follow-up)?

parallel decoding + action chunking reduces latency to 25+ Hz; enables contact-rich, high-frequency manipulation

References​

  • OpenVLA: An Open-Source Vision-Language-Action Model — Kim et al. (2024). arXiv 2406.09246

← Previous: 5.2 π0 — Diffusion-Based Whole-Body Control · Next: 5.4 Diffusion Policy — Visuomotor Control via Denoising →

Part of Module 5: Foundation Models & VLA Architecture.